Image Recognition Model Training via Frequency Domain Desensitization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In federated learning methods for image recognition, data owners face the challenge of information leakage when sharing gradient or weight information, which compromises data privacy during image recognition model training.
Innovation Solution
The method involves performing data desensitization processing on image data using frequency domain transforms and image mixing through Mixup data augmentation, allowing local training of image recognition models without reconstructing original data, thus protecting privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If gradient information or weight information is shared for aggregation in federated learning, then model training performance is improved, but data privacy is compromised due to information leakage
Solution Approach 1:
The patent introduces an intermediary transformation process (frequency domain transform and image mixing) between the original image data and the shared gradient/weight information. This intermediary layer prevents direct reconstruction of original data from shared model updates, thus protecting data privacy while still enabling effective model training through federated learning aggregation.
Solution Approach 2:
The patent applies parameter changes by transforming image data from the spatial domain to the frequency domain, and then applying image mixing operations in this transformed space. These parameter transformations ensure that the shared gradient information contains useful training signals while being resistant to inversion attacks that could reveal original private data.
2Loss of information
If data desensitization processing is applied to protect privacy, then data privacy is improved, but model training effectiveness may deteriorate
Solution Approach 1:
The patent applies preliminary action by performing data desensitization processing (frequency domain transform and image mixing) on the training data before local model training begins. This preprocessing ensures that privacy protection is embedded in the training process from the start, while the transformed data retains sufficient statistical properties for effective model learning.
Data Source
AI summary
Implementations of the present specification provide an image recognition model training method and apparatus. Each first member device performs data desensitization processing on training sample image data based on frequency domain transform to obtain first desensitized image data; provides the first desensitized image data to a hyperparameter selection model to select a first hyperparameter; performs image mixing processing on the first desensitized image data based on data augmentation by using the first hyperparameter to obtain second desensitized image data and second label data that is label mixing processed and corresponding to the second desensitized image data; and then trains an image recognition model by using the second desensitized image data and the second label data. A second member device updates the image recognition model by using a model training result received from each first member device.